After deploying a financial assistant agent, users report occasional inconsistencies in how transactions are categorized.

The runtime should therefore be built around a memory hierarchy that balances retrieval latency, relevance, privacy, and context-window cost. This is a lifecycle problem, not a wording problem, and Option D gives the team a controllable lifecycle for the agent behavior. Transaction categorization depends on tool inputs and outputs. Before retraining, inspect recent traces to see whether the model received incorrect or incomplete structured data. For a production build, memory is an orchestration concern as much as a model concern, because the agent must decide what to keep, retrieve, and forget. The selected option specifically D states “Review tool call inputs and outputs in recent session logs”, which matches the operational requirement rather than a superficial wording match. The rejected options are weaker because sending full history every turn inflates latency and cost, while stateless prompts lose unresolved tasks, user preferences, and multi-step plan continuity. The answer is therefore about engineered control planes, not simply model capability.